Business Problem & Solution

Inbound guest-support email is the slow lane of hospitality. A message arrives in any of a dozen languages, with no category, no priority, and no link to the guest or their booking. Before an agent can actually help, they pay a "triage tax": read it, guess the language, decide whether it's a refund or a complaint or a facilities question, judge how urgent it is, then hunt across other systems for the booking reference, the property, and the loyalty tier. Only then do they start writing — often in a language they don't speak fluently.

Support Email Intelligence collapses that tax into one operator pass. A single metered pipeline cleans the email, detects its language, classifies it, enriches it from the shared data layer, and drafts a reply in the guest's language — leaving a human to do the part that needs judgment: edit, send, or route.

It proves the whole triage loop works on one shared spine: the email, its classification, the enrichment join, the drafted reply, and the cost of the AI run all flow through the same typed adapter and the same event bus the rest of the ecosystem uses. Route an email and the ticket appears in #16's inbox; send a reply and a confirmation lands in #35's sent viewer — no app holds a private copy of the conversation.

It also proves the honesty model: every metered run records a cost-ledger row with its mode and model, owner runs are canonical while a viewer's run is ephemeral, and all of that instrumentation lives only in the inspector — never on the working surface.

The pipeline is honestly labelled simulated. In this prototype the classification is deterministic and rule-based (multilingual keyword + script signals), the drafts are curated per-category templates framed in the detected language, and there is no API key — it runs OSS-only at $0. The adapter contract (`classifyEmail` / `buildDraftReply`) is the same surface a real dual-mode model run is swapped behind in Stage-2, so the UI never changes.

The synthetic mailbox is ~16 invented emails spanning English, French, German, Spanish, and Japanese, joined to synthetic guests and bookings. No real customers, PII, or copyrighted content; any figures shown are representative and labelled. Outbound replies are simulated — they appear in the in-app sent viewer, never over real SMTP.